The Timing of Lexical Memory Retrievals in Language Production

Proceedings of NAACL-HLT 2018, pages 2017–2027 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics The Timing of Lexical Memory Retrievals in Language Production Jeremy R. Cole and David Reitter College of Information Sciences and Technology The Pennsylvania State University University Park, PA jrcole,reitter@psu.edu Abstract This paper explores the time course of lexi- cal memory retrieval by modeling fluent lan- guage production. The duration of retrievals is predicted using the ACT-R cognitive archi- tecture. In a large-scale observational study of a spoken corpus, we find that language pro- duction at a time point preceding a word is sped up or slowed down depending on activa- tion of that word. This computational analy- sis has consequences for the theoretical model of language production. The results point to interference between lexical and phonological stages as well as a quantifiable buffer for lexi- cal information that opens up the possibility of non-sequential retrievals. 1 Introduction Speech varies greatly in fluency, and some of its speed variation can be traced to the utterance spo- ken (Jespersen, 1992). Low-frequency words, for instance, are known to slow down speech (e.g., Bell et al., 2009). Variables correlated with flu- ency give valuable cues to the architecture of the language processing system. However, a model to explain these data has yet to emerge. In this paper, we propose a cognitive model of fluency, in which lexical memory retrievals may explain some of the variability in speech rates. In particular, frequency, context and recent uses to- gether have the potential to quantify retrieval de- lays through activation (Anderson, 1991). Acti- vation, in its most common usage, refers to the way nodes in semantic networks become easier to retrieve after adjacent nodes have been activated, typically through a presentation (Collins and Lof- tus, 1975). In particular, activation makes a direct claim that more highly activated words require less time to retrieve, and vice versa (Anderson, 1983). The language production process as a whole likely requires some amount of sequential process- ing. For instance, the standard model proposes that an idea is generated, lexicalized, grammati- cally and morphologically encoded, and only then phonologically encoded (Bock and Levelt, 2002). Still, most models of language production pre- suppose some amount of planning of output (e.g., Pickering and Garrod, 2013), so we could instead divide language production into planning this out- put and the actual process of outputting. The over- lap and relationship of these processes is not fully understood, but given that most output is likely planned, the scale at which the planning takes place and the amount of time between planned output and the actual process of outputting re- mains unclear. However, if interactions between processes are observed, then we can likewise see when they overlap in time. To summarize, we are suggesting that some of the variance in speech rate is not due to the lin- guistic properties of the words currently or about to be outputted, but the words still in the plan- ning phase. We …

Content truncated to 500 words